1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
|
import unittest
import hypothesis.strategies as st
from hypothesis import given, assume, settings
import numpy as np
import time
import os
from caffe2.python import core, dyndep
import caffe2.python.hypothesis_test_util as hu
dyndep.InitOpsLibrary("@/caffe2/caffe2/contrib/nnpack:nnpack_ops")
np.random.seed(1)
def benchmark(ws, net, warmups=5, iters=100):
for _ in range(warmups):
ws.run(net)
plan = core.Plan("plan")
plan.AddStep(core.ExecutionStep("test-step", net, iters))
before = time.time()
ws.run(plan)
after = time.time()
print("Timing network, time taken per-iteration: {:.6f}ms".format((
after - before) / float(iters) * 1000.0))
return after - before
def has_avx2():
import subprocess
try:
subprocess.check_output(["grep", "avx2", "/proc/cpuinfo"])
return True
except subprocess.CalledProcessError:
# grep exits with rc 1 on no matches
return False
@unittest.skipIf(not has_avx2(), "NNPACK requires AVX2")
class NNPackOpsTest(hu.HypothesisTestCase):
@given(stride=st.integers(1, 3),
pad=st.integers(0, 2),
kernel=st.integers(3, 5),
size=st.integers(5, 10),
input_channels=st.integers(1, 8),
batch_size=st.integers(1, 5),
groups=st.integers(1, 2))
def test_convolution_correctness(self, stride, pad, kernel, size,
input_channels,
batch_size, groups):
input_channels *= groups
output_channels = int(input_channels / groups)
assume(input_channels % groups == 0)
assume(output_channels % groups == 0)
assume(output_channels == input_channels / groups)
assume(stride <= kernel)
if stride != 1:
assume(batch_size == 1)
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
w = np.random.rand(
input_channels, output_channels, kernel, kernel).astype(np.float32)\
- 0.5
b = np.random.rand(output_channels).astype(np.float32) - 0.5
order = "NCHW"
outputs = {}
for engine in ["", "NNPACK"]:
op = core.CreateOperator(
"Conv",
["X", "w", "b"],
["Y"],
stride=stride,
kernel=kernel,
pad=pad,
order=order,
kts="TUPLE",
engine=engine,
group=groups,
)
self.ws.create_blob("X").feed(X)
self.ws.create_blob("w").feed(w)
self.ws.create_blob("b").feed(b)
self.ws.run(op)
outputs[engine] = self.ws.blobs["Y"].fetch()
np.testing.assert_allclose(
outputs[""],
outputs["NNPACK"],
atol=1e-4,
rtol=1e-4)
@given(size=st.sampled_from([6, 8]),
input_channels=st.integers(1, 8),
batch_size=st.integers(1, 5))
def test_max_pool_correctness(self, size, input_channels, batch_size):
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
order = "NCHW"
outputs = {}
# only 2 * 2 stride and 2 * 2 pool is supported in NNPack now
stride = 2
kernel = 2
# The pooling strategy of NNPack is different from caffe2 pooling
pad = 0
for engine in ["", "NNPACK"]:
op = core.CreateOperator(
"MaxPool",
["X"],
["Y"],
stride=stride,
kernel=kernel,
pad=pad,
order=order,
engine=engine,
)
self.ws.create_blob("X").feed(X)
self.ws.run(op)
outputs[engine] = self.ws.blobs["Y"].fetch()
np.testing.assert_allclose(
outputs[""],
outputs["NNPACK"],
atol=1e-4,
rtol=1e-4)
@given(size=st.sampled_from([6, 8]),
input_channels=st.integers(1, 8),
batch_size=st.integers(1, 5))
def test_relu_correctness(self, size, input_channels, batch_size):
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
outputs = {}
for engine in ["", "NNPACK"]:
op = core.CreateOperator(
"Relu",
["X"],
["Y"],
engine=engine,
)
self.ws.create_blob("X").feed(X)
self.ws.run(op)
outputs[engine] = self.ws.blobs["Y"].fetch()
np.testing.assert_allclose(
outputs[""],
outputs["NNPACK"],
atol=1e-4,
rtol=1e-4)
@given(size=st.sampled_from([6, 8]),
input_channels=st.integers(1, 8),
batch_size=st.integers(1, 5),
alpha=st.floats(0, 1))
def test_leaky_relu_correctness(self, size, input_channels, batch_size,
alpha):
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
outputs = {}
for engine in ["", "NNPACK"]:
op = core.CreateOperator(
"LeakyRelu",
["X"],
["Y"],
alpha=alpha,
engine=engine,
)
self.ws.create_blob("X").feed(X)
self.ws.run(op)
outputs[engine] = self.ws.blobs["Y"].fetch()
np.testing.assert_allclose(
outputs[""],
outputs["NNPACK"],
atol=1e-4,
rtol=1e-4)
@settings(deadline=3600)
@unittest.skipIf(not os.environ.get("CAFFE2_BENCHMARK"), "Benchmark")
@given(stride=st.integers(1, 1),
pad=st.integers(0, 2),
kernel=st.sampled_from([3, 5, 7]),
size=st.integers(30, 90),
input_channels=st.sampled_from([3, 64, 256]),
output_channels=st.sampled_from([32, 96, 256]),
batch_size=st.sampled_from([32, 64, 96, 128]))
def test_timings(self, stride, pad, kernel, size,
input_channels, output_channels, batch_size):
assume(stride <= kernel)
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
w = np.random.rand(output_channels, input_channels,
kernel, kernel).astype(np.float32) - 0.5
b = np.random.rand(output_channels).astype(np.float32) - 0.5
order = "NCHW"
times = {}
for engine in ["", "NNPACK"]:
net = core.Net(engine + "_test")
net.Conv(
["X", "W", "b"], "Y",
order=order,
kernel=kernel,
stride=stride,
pad=pad,
kts="TUPLE",
engine=engine,
)
self.ws.create_blob("X").feed(X)
self.ws.create_blob("W").feed(w)
self.ws.create_blob("b").feed(b)
self.ws.run(net)
times[engine] = benchmark(self.ws, net)
print("Speedup for NNPACK: {:.2f}".format(
times[""] / times["NNPACK"]))
@settings(deadline=3600)
@unittest.skipIf(not os.environ.get("CAFFE2_BENCHMARK"), "Benchmark")
@given(size=st.integers(30, 90),
input_channels=st.sampled_from([3, 64, 256]),
batch_size=st.sampled_from([32, 64, 96, 128]))
def test_relu_timings(self, size, input_channels, batch_size):
X = np.random.rand(
batch_size, input_channels, size, size).astype(np.float32) - 0.5
times = {}
for engine in ["", "NNPACK"]:
net = core.Net(engine + "_test")
net.Relu(
["X"],
["Y"],
engine=engine,
)
self.ws.create_blob("X").feed(X)
self.ws.run(net)
times[engine] = benchmark(self.ws, net)
print("Speedup for NNPACK: {:.2f}".format(
times[""] / times["NNPACK"]))
|
